hamedmp / imageflow Goto Github PK

View Code? Open in Web Editor NEW

163.0 163.0 60.0 63 KB

A simple wrapper of TensorFlow for Converting, Importing (and Soon, Training) Images in tensorflow.

License: Apache License 2.0

Python 100.00%

imageflow tensorflow wrapper

imageflow's Introduction

Amplify for Jekyll

A Jekyll html theme in the vague style of Medium.com built using Google AMP

Google's Accelerated Mobile Pages Project (a.k.a. "Google AMP" or Google ⚡) is an open-source project that defines rules for creating websites that load nearly instantly even on mobile devices with slow connections.

Check out a live example of this theme at http://ageitgey.github.io/amplify/ or via Google's CDN.

Screenshot

Why use Google AMP?

There are two reasons to use Google AMP:

First, it's really fast! An often cited claim (by Amazon, Yahoo, Walmart and others) is that every extra 100ms improvement in page load time increases incremental revenue by up to 1%. Your personal blog might not be selling anything, but why settle for a slow page and risk losing readers?

Second, Google might feature your AMP page in Search Results! Google gives preferential treatment to AMP pages on Mobile Search. When it displays your page in the AMP search results widget, it will even serve your page through it's own CDN to make the page load even faster. It's similar to how Facebook Instant Articles works on the Facebook platform.

How fast is this theme?

To get a general idea of how this theme performs, let's compare this page hosted on Github vs. another static page hosted on Github. We can use https://facebook.github.io/react/ as a comparison page. I've also included https://jekyllrb.com/ as another point of comparison (it's also hosted on Github).

Of course our page and these other pages have different layouts. But the main point is that they are typical static sites hosted on Github and are well-designed. So it should give us a rough idea of how other typical pages might perform. I'm not suggesting anything is wrong with these other pages. They are actually pretty fast!

If you are on a fast connection, all these pages load about the same speed but our page renders the main content much faster:

First page visit with no throttling

Page	DOMContentReady	Load
https://facebook.github.io/react/	1.7s	1.89s
https://jekyllrb.com/	500ms	909ms
https://ageitgey.github.io/amplify/2016/03/08/example-post.html	61ms!	1.06s

Second page visit with no throttling

Page	DOMContentReady	Load
https://facebook.github.io/react/	1.08s	1.33s
https://jekyllrb.com/	212ms	486ms
https://ageitgey.github.io/amplify/2016/03/08/example-post.html	66ms!	1.03s

You'll see the main content render much faster because AMP doesn't allow anything in your page that would block the page from rendering after the initial HTML loads. This means no external css, no custom js, etc.

Here's how this looks to the user (as rendered by WebPageTest):

You can get sometimes get even faster speeds when your page is served via Google's AMP CDN. But that's not always true depending the randomness of the internet and where you are connecting from.

So there's some tiny benefit on a 100mbs wired connection. But optimization is much more important on a slow, high-latency mobile connection (i.e. most actual internet users in 2016). Let's try loading the page using the "Regular 2G (250kb/s, 300ms RT)" throttling setting in Chrome Dev Tools:

First page visit with "Regular 2G" throttling

Page	DOMContentReady	Load
https://facebook.github.io/react/	28.50s	29.39s
https://jekyllrb.com/	1.75s	7.03s
https://ageitgey.github.io/amplify/2016/03/08/example-post.html	530ms!	5.07s

Second page visit with "Regular 2G" throttling

Page	DOMContentReady	Load
https://facebook.github.io/react/	2.02s	2.55s
https://jekyllrb.com/	392ms	791ms
https://ageitgey.github.io/amplify/2016/03/08/example-post.html	385ms!	1.64s

Even a horribly slow connection with high latency, the user will still see a page render in half a second. That's great! The difference between 385ms and 28s is the difference between someone reading your blog is skipping your blog.

But notice that the Jekyll homepage still performs well on the second page load. Google AMP gives you a nice set of rules for making fast pages, but of course it isn't required to make a fast page.

Getting Started

To use this theme, it's just like using any other Jekyll template:

Step 1: Install Jekyll

On windows If on windows you will need the ruby devkit available here: rubyinstaller.

Step 2: Clone this repo to your computer

git clone [email protected]:ageitgey/amplify.git

Step 3: Run gem install bundler; bundle install inside the new /amplify/ folder that was just created to install the required ruby dependencies.

Step 4: Tweak _config.yml.

Just fill in everything in the # Site settings section. You'll want to set your site's title, your name, your twitter username, etc.

Step 5: Run bundle exec jekyll serve and then open http://localhost:4000/ to see your site!

Step 6: Publish your site just like any other Jekyll site.

Google AMP Limitations

Google AMP sets many strict limits on what you can include in your web pages. A few of these are worth talking about:

Limitation: All CSS must be inline (no external css files).

Because of this, the main css file for this site is in _includes/styles.scss instead of in the normal css/ Jekyll folder. This css file is inlined into the header of every page via the special scssify filter in _includes/head.html.

Limitation: Size all resources statically

Every image you include in your page must have a height and width. This also applies to other things like embedding videos or other resources. Check below for more details.

Writing Posts with Google AMP

Writing posts works just like it does normally in Jekyll except when you want to include extra resources likes pictures, videos, embedded Twitter posts, etc.

Google AMP has it's own set of special html tags for including content. You should use these instead of normal Markdown or HTML tags.

The two you are are most likely to need are <amp-img> and <amp-youtube>:

Images in your posts

<amp-img width="600" height="300" layout="responsive" src="/assets/images/your_picture.jpg"></amp-img>

Youtube Videos in your posts

<amp-youtube data-videoid="lBTCB7yLs8Y" layout="responsive" width="480" height="270"></amp-youtube>

Embedding other types of content

The AMP Project provides helpers for many other types of content like audio, ads, Google Analytics, etc.

Built-in AMP tags:
https://github.com/ampproject/amphtml/blob/master/builtins/README.md
Extended AMP tags:
https://github.com/ampproject/amphtml/blob/master/extensions/README.md

Validating your page with Google AMP

Google AMP adds built-in validation logic to make sure your pages follow all the rules so they render as fast as possible.

To check your page, just add #development=1 to any url on your site and then check the javascript console in your browser.

http://localhost:4000/#development=1

You will either see a success message:

Powered by AMP ⚡ HTML – Version 1457112743399
AMP validation successful.

Or you will see a list of errors to fix:

Powered by AMP ⚡ HTML – Version 1457112743399
AMP validation had errors:
The attribute 'style' may not appear in tag 'span'
The attribute 'style' may not appear in tag 'div'

Enabling Google Analytics

This theme supports simple page tracking with Google Analytics.

To enable analytics :

Set your property ID in _config.yml
Uncomment the analytics include in _layouts/default.html.
Uncomment the analytics script in _includes/head.html.

If you wish to track custom events or want to send custom data, please refer to the documentation.

Making Google serve your page

Google will cache valid AMP pages if you link to them with one of these urls:

https://cdn.ampproject.org/c/s/<your page url here>

Or:

https://amp.gstatic.com/c/s/<your page url here>

But keep in mind these two limitations:

The caches don't refresh that often. So don't view these urls until your page is done!
Remove /s from both urls if your page isn't served over https://.

Required Schema Data

To actually get your page featured in Google search results, it needs to include a http://schema.org NewsArticle schema. See _includes/metadata.json for the version generated by default. You might want to tweak it.

Credits

This theme is inspired by Mediator by Dirk Fabisch. I used some of the css and html from that theme as a starting point. Thanks!

License

MIT. See LICENSE file in repo.

imageflow's People

Contributors

Stargazers

Watchers

imageflow's Issues

my_cifar_train.py variables initialisation issue

from: init = tf.initialize_all_variables()
to: init = tf.group(tf.initialize_all_variables(), tf.initialize_local_variables())
as stated in: http://stackoverflow.com/questions/38136081/setting-num-epochs-on-tf-train-string-input-producer-produces-an-error

make a prediction with new image?

Hi HamedMP,

I have been studying this library and found is very amazing.
I trained my own model and I want to make prediction to having a predicted label with a new image.
Not sure how to do it, may you share your knowledge around this? Thanks.

Regards,
Deeperic

tensorflow tf.image.resize_images raise ValueError

I used ImageFlow to create a tensorflow tfrecords file on my own images. My images are of the same size 223x334x3. I tried the cifar example on my own images. Because the cifar convnet need image of 32x32 size so i have to resize it.

My function to read tfrecords file , decode images and resize them as follow:

def read_my_file_format(filename_queue):
reader = tf.TFRecordReader()
key, serialized_example = reader.read(filename_queue)

features = tf.parse_single_example(
serialized_example,
features={
  'image_raw': tf.FixedLenFeature([], tf.string),
  'label': tf.FixedLenFeature([], tf.int64)
})

image = tf.decode_raw(features['image_raw'], tf.uint8)
image = tf.cast(image, tf.float32)
label = tf.cast(features['label'], tf.int32)

image = tf.reshape(image, [223446])
image.set_shape(223446)
print 'get shape: ',image.get_shape()

print 'image: ',image.get_shape()
print 'label: ',label.get_shape()

image = tf.image.resize_images(image, 32,32)
image = tf.reshape(image,[32,32,3])

#processed_example = some_processing(example)

processed_example = image
return processed_example, label

But when i call it, i got the following error:

Last executed 2016-07-22 23:58:40 in 71ms
get shape:  (223446,)
image:  (223446,)
label:  ()
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-32-44cda3e31e6a> in <module>()
----> 1 tf.app.run()

/home/guanwanxian/anaconda2/envs/theano/lib/python2.7/site-packages/tensorflow/python/platform/app.pyc in run(main)
     28   f._parse_flags()
     29   main = main or sys.modules['__main__'].main
---> 30   sys.exit(main(sys.argv))

<ipython-input-7-2788f377abf1> in main(argv)
      1 def main(argv=None):
----> 2     train()

<ipython-input-31-4c9a489f07f7> in train(re_train, continue_from_pre)
      8         # Get images and labels for CIFAR-10.
      9         # images, labels = my_input.inputs()
---> 10         images, labels = input_pipeline(['../data/trainV2.tfrecords'],batch_size=128)
     11         val_images, val_labels = input_pipeline(['../data/testV2.tfrecords'],batch_size=128)
     12 

<ipython-input-3-ce4e93d1c196> in input_pipeline(filenames, batch_size, num_epochs)
      2     filename_queue = tf.train.string_input_producer(
      3                      filenames, num_epochs=num_epochs, shuffle=True)
----> 4     example, label = read_my_file_format(filename_queue)
      5     # min_after_dequeue defines how big a buffer we will randomly sample
      6     #   from -- bigger means better shuffling but slower start up and more

<ipython-input-30-8be4eeea8a06> in read_my_file_format(filename_queue)
     21     print 'label: ',label.get_shape()
     22 
---> 23     image = tf.image.resize_images(image, 32,32)
     24     image = tf.reshape(image,[32,32,3])
     25 

/home/guanwanxian/anaconda2/envs/theano/lib/python2.7/site-packages/tensorflow/python/ops/image_ops.pyc in resize_images(images, new_height, new_width, method, align_corners)
    629     images = array_ops.expand_dims(images, 0)
    630 
--> 631   _, height, width, depth = _ImageDimensions(images)
    632 
    633   # Handle tensor-valued sizes as well as Python integers.

ValueError: need more than 1 value to unpack

and My code to create tfrecords from images with littlt modification as follow:

def read_image_and_labels_from(path):
'''read images and return image and label array
'''

directories = glob.glob(os.path.join(path,'*'))
class_names = [os.path.basename(directory) for directory in directories]
class_names.sort()
num_classes = len(class_names)

file_paths = glob.glob(os.path.join(path,'*/*'))
file_paths = sorted(file_paths,
       key=lambda filename:os.path.basename(filename).split('.')[0])

images = []
labels = []
shapes = []
shape_set = set()
for filename in file_paths:
    im = Image.open(filename)
    arr = np.asarray(im, np.uint8)
    images.append(arr)

    im_shape = arr.shape
    shapes.append(im_shape) # tuple of 3
    if im_shape not in shape_set:
        shape_set.add(im_shape)
    # image_name = os.path.basename(filename).split('.')[0]

    class_name = os.path.basename(os.path.dirname(filename))
    label_num = class_names.index(class_name)
    labels.append(np.asarray(label_num, np.uint32))

images = np.array(images)
labels = np.array(labels)
shapes = np.array(shapes)
print images.shape, labels.shape, shapes.shape
print 'shape set: ', shape_set
return images, labels, shapes


def convert_images_to_tfrecords_file(images, labels, shapes, directorys, name):
num_examples = labels.shape[0]
print('labels shape is: ', labels.shape[0])
if num_examples != images.shape[0] != labels.shape[0]:
    raise ValueError("Images size %d does not match label size %d." %
                                     (images.shape[0], num_examples))

filename = os.path.join(directorys, name + '.tfrecords')
print('Writing', filename)
writer = tf.python_io.TFRecordWriter(filename)
for index in range(num_examples):
    image_raw = images[index].tostring()
    image_shape = shapes[index]
    example = tf.train.Example(features=tf.train.Features(feature={
            'height': _int64_feature(image_shape[0]),
            'width': _int64_feature(image_shape[1]),
            'depth': _int64_feature(image_shape[2]),
            'label': _int64_feature(int(labels[index])),
            'image_raw': _bytes_feature(image_raw)}))
    writer.write(example.SerializeToString())

So what's wrong with my code?

per_image_whitening function does not exist

The function tensorflow.image.per_image_whitening (used here) no longer exists.

It has been renamed in TensorFlow 0.12 to tensorflow.image.per_image_standardization. See TensorFlow 0.12 release notes

IndexError: tuple index out of range in convert_to

Dear @HamedMP ,
I use files in example folder for reading and convert images to tfrecords file. but, I have some errors:
I using grayscale transform, so my images are 2D array.
first, I had a "too many indices" error in line: validation_images = train_images[:FLAGS.validation_size, :, :, :]
and with changing [:FLAGS.validation_size, :, :, :] to [:FLAGS.validation_size] this worked well.
besides that, my other issue is "IndexError: tuple index out of range" in convert_to function( in line "rows = images.shape[1]").
I don't understand, why we are using "rows = images.shape[1]" whereas our images have different size and even more, they have no depth in grayscale version.

best,
MRHajbabaei.

Name conflict - disambiguation?

Hi, I'm the author of https://github.com/imazen/imageflow - another image processing tool. Do you think we should add a disambiguation paragraph on each repo?

Create dataset from scratch

First of all, thank you @HamedMP for the huge work.
In my opinion, to extend ImageFlow functionalities, a nice feature to add is giving the possibility to create a new dataset from scratch.
It could be very usefull to anyone who needs to create his personal cnn for image recognition not base on pre-built dataset (i.e. Cifar, MNIST or Imagenet).
What do you think ?

Provide the test example for cifar prediction

Would you provide a test example based ontest.tfrecords?
After the training of the model is finished, I want to get the predicting labels for every testing image and save them in a file.

Multilabel classification example

Would it be possible to also include an example of how to create the TFRecord with multiple labels? I've been working on this, but am stuck--not sure if I'm not interpreting the example incorrectly or what the issue is.

In def convert_to, I've changed 'label': _int64_feature(int(labels[index])) to 'label': labels[index,:], where labels is an array of n by 5--n examples with 5 different labels representing 5 different attributes of a possible image. To create the labels array, I use:

labels[i,:] = [_int64_feature(int(l1)), _int64_feature(int(l2)), _int64_feature(int(l3)), _int64_feature(int(l4)), _int64_feature(int(l5))]

where I loop through each line of my text file, parse it for the 5 specific numbers I need (read in as strings), cast them as ints, and then pass them to the _int64_feature function.

However, I get errors with only _int64_feature(int(l5)), where it says that TypeError: float() argument must be a string or a number. When I wasn't getting that error, using the convert_to function led to this error: TypeError: Parameter to MergeFrom() must be instance of same class: expected Feature got ndarray.

Apologies if this isn't the correct place to ask this question, I just figured that it'd be helpful for anyone else who's working on a similar thing to be able to refer to this issue to resolve their errors as well.

ImportError: No module named 'reader'

I'm currently running Jupyter notebook using tensorflow-notebook . When I try import imageflow, I get the following:

---------------------------------------------------------------------------
ImportError                               Traceback (most recent call last)
<ipython-input-5-f680d40931f4> in <module>()
----> 1 from imageflow import *

/opt/conda/lib/python3.5/site-packages/imageflow/__init__.py in <module>()
     27 
     28 from imageflow import *
---> 29 from reader import read_and_decode
     30 
     31 __author__ = 'HANEL'

ImportError: No module named 'reader'

Might-be-relevant-software versions:
Python - 3.5.2
Jupyter - 4.1.1
numpy - 1.10.4
Pillow - 3.2.0
tensorflow - 0.9.0
imageflow - 0.0.2

EDIT: Looks like it's a problem with Jupyter notebook, as everything works fine from the command line, so I'll close this.

reader.py missing

It seems that reader.py doesn't get installed when you do pip install imageflow.

edit: I looked in my package directory and did see the reader.py file and util.py files.. but for some reason It throws an error when I attempt to import imageflow. Says reader is missing... was able to fix it by puting the files in the local directory of the py file I am working with.

Anaconda

How can I install ImageFlow in Anaconda?

Image flow example: Receive only a portion of tfrecords images

Dear @HamedMP

I use your code in example folder and tried to train my network with that. However, when I start the training process I receive only a portion of my whole images in my tfrecords file (actually there is a formula for that! Number_of_images = 172*(FLAGS.num_epochs/FLAGS.batch_size), I guess).
I know that coord.should_stop() try to stop network from training if a thread should stop. But, I cannot realize why that try to stop network very fast. despite the fact that, my epoch number not yet completed?

Another question: what does FLAGS.num_epochs do in this code? (I don't realize difference of FLAGS.num_epochs in your code and epoch_numer=all_data_size/batch_size)

thank you very much.
mohammad reza